When AI can't find you, it invents you

Large language models complete patterns. Ask about a brand with thin source coverage and the model does not say "I don't know" — it generalizes from category patterns, similar names, and stale training data. The outputs we routinely catch in audits: wrong founding years, invented product lines, merged-up competitors (two companies fused into one entity), stale pricing, and sentiment frozen at a years-old controversy. None of these are malice; all of them are downstream of a sparse, inconsistent source graph.

The three sub-scales

  • Fact accuracy (50% of the score). Extract every factual claim engines make about you across the prompt set — founding date, HQ, products, certifications, pricing, key people — and verify each against ground truth. Score = correct claims / total claims.
  • Sentiment valence (30%). Classify the framing of each brand mention: recommending, neutral-descriptive, hedged, or negative. Track the wording — "a solid budget option" and "the category leader" are both positive and miles apart.
  • Cross-engine consistency (20%). Do ChatGPT, Gemini, Perplexity and AI Overviews tell the same story? Divergence reveals engine-specific source problems — e.g. a negative Reddit thread feeding one engine's retrieval but not the others'.

Correcting wrong facts: the source-first method

  1. Trace the claim to its source. Ask the engine follow-ups ("where did you read that?") and inspect cited URLs. Most hallucinated facts trace to one stale directory entry, one old article, or an ambiguous passage on your own site.
  2. Fix at the source, not in the prompt. Correct the directory, update the article, clarify your page. Editing your own site alone leaves the contradicting source in the retrieval pool.
  3. Add corroboration. One corrected source against three stale ones loses. Publish the corrected fact in 3+ independent places engines read.
  4. Re-test on a schedule. Models refresh unevenly — a fix can show in Perplexity (live retrieval) within days and take weeks to reach training-data-dependent answers.
Negative-sentiment emergencies: when an engine answers "is [brand] legit?" with hesitation, the cause is almost always a specific, findable set of sources. Panic-posting new content without removing the contradicting sources is the most common wasted spend we see. Trace → fix → corroborate → re-test.

Frequently asked questions

Can I just ask ChatGPT to forget wrong information about my brand?
No. There is no self-service correction channel that edits model knowledge directly. Correction works through the source graph: fix the underlying sources engines retrieve and cite, and answers follow. OpenAI and others accept legal takedowns in narrow cases, but factual correction is a source-management problem.
How do I find out what AI says about my brand?
Run a structured prompt set — 'what is [brand]', 'is [brand] reliable', '[brand] vs [competitor]', '[brand] reviews', '[brand] pricing' — across all major engines, and log every claim. The GEO Checker automates this and produces the fact-accuracy and consistency report.
How long does it take to fix a hallucinated fact?
For engines with live retrieval (Perplexity, ChatGPT search), days to weeks after source correction. For answers relying more on training data, weeks to a model-refresh cycle. Budget 30–90 days and re-test monthly.